Installation

Currently Torsten is based on a forked version of Stan and hosted on GitHub

• https://github.com/metrumresearchgroup/Torsten

The latest v0.89rc is compatible with Stan v2.27.0. Torsten can be accessed from command line for cmdstan interface and cmdstanr (https://mc-stan.org/cmdstanr/) for R interface. It requires a modern C++11 compiler as well as a Make utility. See (Team 2020) for details of installation and required toolchain. In particular, we recommend the folowing versions of C++ compilers:

• Linux: g++ >=7.5 or clang >=8.0,
• macOS: the XCode version of clang,
• Windows: g++ 8.1 (available with RTools 4.0).

On windows, the Make utility mingw32-make can be installed as part of RTools.

Command line interface

The command line interface cmdstan is available along with Torsten and can be found at Torsten/cmdstan.

After installation, one can use the following command to build a Torsten model model_name in model_path

cd Torsten/cmdstan
make model_path/model_name # replace "make" with "mingw32-make" on Windows platform

R interface

After installing cmdstanr from https://mc-stan.org/cmdstanr/, use the following command to set path

cmdstanr::set_cmdstan_path("Torsten/cmdstan")

Then one can follow https://mc-stan.org/cmdstanr/articles/cmdstanr.html to compile and run Torsten models.

MPI support

Torsten’s MPI support is of a different flavour than reduce_sum found in Stan. To be able to utilize MPI parallelisation, one first needs to ensure an MPI library such as

• https://www.mpich.org/downloads/
• https://www.open-mpi.org/software/ompi/

is available. Torsen’s implementation is tested on both MPICH and OpenMPI.

To use MPI-supported population/group solvers, add/edit make/local

TORSTEN_MPI=1

# path to MPI headers
CXXFLAGS += -isystem /usr/local/include
# if you are using Metrum's metworx platform, add MPICH3's
# headers with
# CXXFLAGS += -isystem /usr/local/mpich3/include

Note that currently TORSTEN_MPI and STAN_MPI flags conflict on processes management and cannot be used in a same Stan model, and MPI support is only available through cmdstan interface.

Testing

Models in example-models directory are for tutorial and demonstration. The following shows how to build and run the two-compartment model using cmdstanr, and use bayesplot to examine posterior density of CL.

library("cmdstanr")
set_cmdstan_path("Torsten/cmdstan")
file.dir <- file.path("Torsten", "example-models", "pk2cpt")
file  <- file.path(file.dir, "pk2cpt.stan")
model <- cmdstan_model(file)
fit <- model$sample(data = file.path(file.dir, "pk2cpt.data.R"),
                    init = file.path(file.dir, "pk2cpt.init.R"),
                    seed = 123,
                    chains = 4,
                    parallel_chains = 2,
                    refresh = 500)
bayesplot::mcmc_dens_overlay(fit$draws("CL"))

Bibliography

Team, Stan Development. 2020. CmdStan User’s Guide. https://mc-stan.org/docs/2%5F26/cmdstan-guide/index.html.